Contactless palmprint identification using CompNet on the IITD and SMPD databases

Contactless palmprint recognition is attractive for mobile biometrics, yet how well a learned representation transfers between acquisition settings is rarely measured. Using CompNet - a compact network of learnable Gabor filters and competitive coding — we apply one unified protocol to two contrasting databases: the constrained near-infrared IITD and the unconstrained smartphone SMPD. Within each, CompNet reaches 98.70% Rank-1 accuracy at an equal-error rate below 1.4%. Across databases the transfer is asymmetric: mobile-trained features generalize to controlled data (91.10%), whereas the reverse collapses (40.03%), so generalization flows from harder to easier domains.